Identity information based on human magnetocardiography signals

TL;DR

The paper addresses biometric identification using non-contact, physiological signals by leveraging magnetocardiography (MCG) signals captured with optically pumped magnetometers (OPMs). It converts spatially distributed chest-MCG data into time-frequency representations via a wavelet transform and uses a CNN with 4-channel inputs (constructed from 2×2 neighboring signals) to classify individuals. The authors report a macro F1-score of $97.04\%$ for five-subject identification and an accuracy of $97.03\%$ on a held-out test set, with robustness to moderate noise, all achieved without a magnetically shielded room. This approach demonstrates the potential of room-temperature, non-contact MCG-based identity verification for personalized healthcare and security, while acknowledging limitations related to confounding health signals and the need for broader validation across populations and noise conditions.

Abstract

We have developed an individual identification system based on magnetocardiography (MCG) signals captured using optically pumped magnetometers (OPMs). Our system utilizes pattern recognition to analyze the signals obtained at different positions on the body, by scanning the matrices composed of MCG signals with a 2*2 window. In order to make use of the spatial information of MCG signals, we transform the signals from adjacent small areas into four channels of a dataset. We further transform the data into time-frequency matrices using wavelet transforms and employ a convolutional neural network (CNN) for classification. As a result, our system achieves an accuracy rate of 97.04% in identifying individuals. This finding indicates that the MCG signal holds potential for use in individual identification systems, offering a valuable tool for personalized healthcare management.

Figures (5)

• Figure 1: The coordinates of all the points for the measurement. From Y1 to Y7 is aligned along the spine with adjacent rows 5 cm apart and row Y1 aligned with the clavicle. From X1 to X7 is arranged perpendicular to the spine with a distance of 5 cm between adjacent columns. The midpoint of the interval between X3 and X4 is aligned with the intersection of the two clavicles. MCG signals at 49 positions are obtained based on the $7\times 7$ measurement matrix. The denoised MCG diagrams are placed on the corresponding coordinates to show the cardiac magnetic field in front of the chest cavity. The box represents the position relationship between different data channels when MCG signals are converted into deep learning data sets.
• Figure 2: The photograph and schematic diagram of our MCG system. It consists of two Bell-Bloom OPMs as a gradiometer and a set of two-layer 3D Helmholtz coils. The gradiometer is applied to sense the cardiac magnetic field, at room temperature and in natural magnetic environments. The coils adjusts the direction and strength of the bias magnetic field and keeps the magnetic field around the vapor cell to a setpoint.
• Figure 3: MCG signal was measured at the point (X4, Y4) in an environmental magnetic field of 47000 nT. Since the IFN is significantly larger than the intrinsic body activity signal, some high-frequency noise still exists in the MCG signal after filtering. (a) About 2 seconds of MCG signal after denoising. We present the magnetic induction intensity obtained by the probe in pT, and the time elapsed since the measurement began in second, respectively. A positive value and a negative value indicate that the direction of the magnetic induction intensity is perpendicular to the thoracic plane upward and downward, respectively. The peak of the R-wave is about 20 pT, and the peak of the S-wave is about -5 pT. (b) The power spectral density of MCG data after denoising, power spectral density is frequency. IFN still has an effect, but the intrinsic body activity signal is the major component. We present the frequency in Hz and the power spectral density corresponding to the frequency in pT over square Hz, respectively.
• Figure 4: The sensitivity curve in double-classification. Changing the classification threshold resulted in a less difference in accuracy, indicating that most of the test data sets are significantly classified.
• Figure 5: The performance of the trained module in classification when two types of noise are added simultaneously. We present the Gaussian noise intensity and the random noise intensity, respectively. The intensity of noise is defined as the bottom of the SNR. The real line and the dashed line represent the maximum acceptable noise threshold when the requirement for recognition accuracy is above 95% and 85%, respectively. The model is more robust to the Gaussian noise following a normal distribution than the random noise.